Research

1. Bone Water as a Window into Bone Matrix Health 

We read bone through its water. Bone is not simply mineral. Water bound to collagen and mineral is part of how bone is organized as a material, and shifts in that water reveal changes that density alone misses. This is not an academic point. Many people who go on to fracture have unremarkable bone density, and the water signal offers a way to see the risk earlier, before it registers as bone loss.

Our laboratory develops and applies imaging and spectroscopic approaches to understand what bone water tells us about the tissue around it. We use ultrashort echo time MRI to follow water-related changes in living bone. We use near-infrared spectroscopy and spectroscopic imaging to resolve those changes at higher spatial resolution. Neither signal means much on its own, so we anchor both to destructive ground truth: mechanical testing to failure, thermal and compositional analysis, and direct measurement of the collagen-mineral environment. That is the part that lets a measurement made in living bone stand for the material state of the tissue.
 

Not all bone water is the same. Bound water sits with collagen and the organic matrix. Pore water fills the vascular and cellular spaces. As bone quality deteriorates, water moves from the bound compartment to the pore compartment, and this redistribution often shows up before conventional density does. It is one reason bone can change materially while standard skeletal measures stay quiet, which is the question that drives much of our work.

We study how aging, hormonal and metabolic disturbance, disease, and treatment alter the collagen-mineral-water environment, and what those changes mean for skeletal function and adaptation. The work spans models of chronic kidney disease-mineral bone disorder, X-linked hypophosphatemia, and estrogen loss. Across all of them we connect water-based signals to composition, to mechanics, and to the cells that build and maintain bone. Bound water is not the mechanism. It is an in vivo-accessible window into the broader state of the bone.

2. Radiomics and Machine Learning for Bone and Muscle Health 

Reading soft tissue, not just bone. Osteoporosis is diagnosed from bone density, but the tissue around the bone carries information too. In work led by Mohseu Subah, we built an automated pipeline that segments HR-pQCT scans into cortical and trabecular bone of the tibia and fibula along with skin, muscle, and fat, then we extracted radiomic features from each region. Features from muscle and tendon separated people with and without DXA-defined osteoporosis better than features taken from bone. Adding these soft tissue features to the usual biological, DXA, and HR-pQCT measures raised patient-level discrimination (AUROC) from 0.79 to 0.88. The result points to muscle quality as a readable part of skeletal fragility that standard analysis leaves on the table.

Grounding radiomic features in bone biology. Radiomic texture features are easy to compute and hard to interpret. To connect them to something physical, Youngjun Lee led a cadaveric study that stripped bone of its individual building blocks one at a time, mineral, then collagen, then free and bound water, imaging each sample along the way. Every component left its own reproducible signature in the texture features. Mineral loss drove the largest change in radiodensity, while collagen and water each produced distinct and measurable effects. This gives us a compositional basis for saying what a radiomic feature responds to, which matters when the same features are used to flag disease.

Earlier signals of disease in chronic kidney disease. Chronic kidney disease weakens bone in ways DXA can miss. Applying radiomics to HR-pQCT scans of cortical bone, we can separate CKD bone from healthy controls and pick up textural changes before they register as density loss. We are extending the same approach to track muscle quality alongside bone and to follow how tissue changes after kidney transplant.

Segmentation that scales across sites. Large-scale radiomics only works if the underlying segmentation holds up on data from different scanners. A self-configuring network trained at a single institution held up across seven other sites in the FAMuSS muscle MRI cohort without any site-specific tuning. We treat this as the foundation that supports radiomics rather than the finding itself.

3. Motion Correction and Automated Quality Grading in High-Resolution Imaging 

HR-pQCT resolves bone microstructure at about 61 micrometers, which is also what makes it fragile. A small movement during the two-minute scan leaves streaks and blurring that can render the microarchitecture unusable, and the standard fix of rescanning is not always possible for children, older adults, or people with tremors. Correcting that motion after the fact is the harder problem, because supervised models need matched pairs of clean and corrupted images and those pairs do not exist for real patients. You cannot ask someone to hold still and move at the same time.

Our approach is to build the reference data the models need. We model patient motion as rotation and translation of the imaged limb and corrupt the raw sinogram to recreate motion artifacts from clean scans. The simulated streaking matches what we see in genuinely motion-corrupted data, which gives us the paired clean and corrupted images that supervised correction depends on (Sadik et al., IEEE TRPMS 2025). HR-MoCo47K is the first publicly available HR-pQCT motion correction dataset, pairing 38,472 simulated images from 229 participants with 9,072 real motion-corrupted images from another 54. We use a dual CycleGAN domain adaptation framework to bring the simulated and real domains together, since simulated data does not transfer cleanly to real scans on its own (Sadik et al., CVPR Workshops 2026). On severely degraded scans the models recover image quality, though restoring true microstructure without inventing detail is still an open problem.

Correction is only trustworthy if we can measure the motion in the first place. The current visual grading is manual and disagrees with itself across operators, so we built an automated grader that also explains itself. It pairs a frozen vision model (DINOv2) with a biomedical language model (PubMedBERT) so predictions stay tied to the same written descriptors human graders use, reaching weighted F1 near 0.94 for the distal radius and 0.92 for the distal tibia on binary scoring. Beyond the score, the model's attention lands on the cortical breaks and streaks a trained operator would flag.

We are now carrying the same idea beyond bone, extending the motion simulation and correction work to lung MRI reconstructed from PETALUTE radial acquisitions, where breathing and bulk motion create a related problem and clean reference images are hard to come by..